Gains:
- RFM (Recency-Frequency-Monetary) ve davranışsal segmentleri yapay zeka desteğiyle tanımlayıp yorumlayabilme
- Her segment için kişiselleştirilmiş iletişim ve sadakat teklifi taslağı üretebilme
- Segmentasyon ve kişiselleştirmede ayrımcılık, gizlilik ve açık rıza sınırlarını gözetebilme
Not all customers are the same. Some come every week and spend a lot; some come once and disappear; Some used to be loyal but don't stop by anymore. Sending the same message and the same offer to all of them is a waste of money and does not yield results. Müşteri segmentasyonu, müşterileri anlamlı gruplara ayırıp her gruba uygun yaklaşımı kurmaktır. In this unit, you will use artificial intelligence as an assistant that extracts segments from customer data and produces a communication draft suitable for each segment. Kritik sınır: segmentasyon güç verir ama bu güç ayrımcılığa ve gizlilik ihlaline dönüşmemelidir.
Segmentation: It is the process of dividing customers into groups based on similarity of behavior, values or needs.
Personalization: Providing each segment (or person) with content, offers or experiences that suit their interests and needs.
RFM: the most practical segmentation
The most used method in retail is RFM. It looks at three simple signals:
- R (Recency): When did the customer last shop? Recent = more active.
- F (Frequency): How many times did he shop in a certain period? High = more connected.
- M (Monetary): How much did he spend in total? Higher = more valuable.
Each customer is given a score on these three dimensions (e.g. 1-5), and the combination of scores determines the segment. Yapay zeka bu skorlamayı ve segment adlandırmayı hızlıca yapar.
segment
RFM profile
Approach
Champions
High R, F, M
Reward, early access, make ambassadors
loyal customers
High F, mid M
Up-selling, loyalty advantage
those at risk
Low R, formerly high F/M
win back campaign
new customers
High R, low F
Hoş geldin akışı, ikinci alışverişi teşvik
those who fall asleep
Low R, low F
Low cost reminder or drop
Tip: Don't send discounts to everyone. The championship discount is a loss of margin (which it would have received anyway); Save the discount for the real at-risk and dormant segment.
Behavioral segments
Beyond RFM, behavior also establishes segments: category preference (buying only baby products), channel (online only), price sensitivity (buying always on sale), seasonality. Artificial intelligence can extract these patterns from the data and suggest groups such as “price hunters,” “premium buyers,” “weekend shoppers.”
The limits of personalization: ethics and privacy
Segmentation is powerful, but never cross three lines:
- No discrimination: The segment cannot be based on protected characteristics (ethnicity, religion, health, sexual orientation); Making indirect inferences from them (e.g. estimating ethnicity by postcode) is also risky.
- Explicit consent: Explicit consent of the customer is usually required to send commercial electronic messages (SMS, e-mail). Without consent, sending is a legal and reputational risk.
- Data minimization: Use the least data needed to establish segments; Stay away from sensitive data.
Caution: Targeting based on sensitive life events such as "This customer may be pregnant, let's recommend a baby product" may disturb the customer and create the perception of a violation of privacy. Be very careful with inferential targeting.
Step by step segmentation
- Prepare anonymous data: Customer code (not name), last shopping date, number of purchases, total spend, category. Remove personal identifiers.
- RFM score: Have the artificial intelligence score 1-5 in the R, F, M dimensions.
- Name segments: Link score combinations into meaningful groups.
- Action map: Draft goals, offers, and messages for each segment.
- Filter consent and ethics: Send only to consenting customers, remove discriminatory targeting, measure results.
mini cases
Case 1 — The right offer to the right segment: A cosmetics chain was sending uniform 15% discount SMS to 40,000 customers per month. When segmented with artificial intelligence RFM, the "champions" (8,500 people) were already receiving regularly. The discount was sent only to 6,200 people at risk, and early access + points were sent to champions. While the campaign cost decreased by 30%, the number of returning risk customers increased by 18%.
Case 2 — Winning back: An online grocery store moves 3,100 formerly loyal customers who haven't shopped in 90+ days into its "at risk" segment. The AI produces a personalized “we miss you + free shipping from the category you have received in the past” message draft. It is sent to 2,400 people with consent; 410 of them shop again within 6 weeks.
Case 3 — Ethical error and correction: A store wants to use the neighborhood information when requesting a "high spending potential" segment from artificial intelligence. The responsible person realizes that this poses indirect discrimination and privacy risks and bases it on actual shopping behavior rather than neighborhood. The segment is both fairer and more accurate.
Weak prompt / Strong prompt
Weak prompt:
Group my customers.
Which data, which logic, which purpose is unclear.
Powerful prompt:
Your role: CRM and segmentation specialist. Data: Below is the anonymous customer table (customer_code, number of last shopping days, number of purchases in 12 months, total spending, most purchased category). No personal information. Task: (1) divide customers into 5 segments with RFM logic and name each segment, (2) give the approximate size of each segment, (3) one-sentence action suggestion for each segment. Rule: Use only behavioral data in the data; Do not make discriminatory inferences from district, name or demographics. Data: [paste]
Copiable prompt templates
1) RFM segmentation
Score the following anonymous customer data with RFM logic (1-5 for R, F, M) and divide it into meaningful segments. Give the size and description of each segment. Present it in table form. Data: [paste]
2) Segment-based communication draft
Write a personalized SMS and an email draft for the segment: [segment description]. Tone fits our brand [tone], proposal [offer].Suggest two different titles. Do not use misleading claims, 40 words maximum.
3) Winback campaign
Draft a win-back message for old loyal customers who haven't shopped in 90+ days. Based on past category preference, offer a low-cost incentive. Give 3 different approaches.
4) Ethics/consent audit
Check the following targeting plan: is there discrimination based on protected characteristics, sensitive life event inference, non-consensual submission or unnecessary use of personal data? Flag risks and suggest alternatives. Plan: [paste]
Common mistakes
- Same message to everyone: Segmentless communication is a waste of money and attention.
- Championship unnecessary discount: Discount to the customer who will already buy it is a loss of margin.
- Sending without consent: Explicit consent is required for SMS/e-mail.
- Segment based on protected characteristics: Discrimination; Indirect inference is also risky.
- Sensitive life event targeting: Creates the perception of a violation of privacy.
- Set a segment once and forget it: Customer behavior changes; Update regularly.
In summary
Segmentasyon, doğru müşteriye doğru yaklaşımı kurmanın yoludur; RFM is the most practical start. Yapay zeka skorlama, segment adlandırma ve mesaj taslağını hızlandırır. Ama açık rıza, veri minimizasyonu ve ayrımcı olmama sınırları pazarlık dışıdır; base personalization on behavioral data, avoid sensitive inference.
Application task
Anonim müşteri verinizden (müşteri kodu, son alışveriş, sıklık, harcama, kategori) bir tablo hazırlayın. "1) RFM segmentasyon" ile 5 segment çıkarın, en değerli ve en riskli segment için "2) Segment bazlı iletişim taslağı" üretin. Planınızı "4) Etik/rıza denetimi"nden geçirin.
checklist
- [ ] I anonymized customer data.
- [ ] I created meaningful segments with RFM.
- [ ] Her segmente uygun teklif ve mesajı ayırdım.
- [ ] I planned to send only to consenting customers.
- [ ] I have avoided protected property and sensitive inference.
- [ ] I made a plan to update the segments regularly.